CNTF: Commonsense, Named Entity and Topical Knowledge Fused Neural Network

This is an implementation of our paper:

Commonsense and Named Entity Aware Knowledge Grounded Dialogue Generation (NAACL 2022).
Deeksha Varshney, Akshara Prabhakar, Asif Ekbal
Link : https://arxiv.org/abs/2205.13928

Overview

The code and sample data for our work is organized as:

• source/ contains the main model scripts
• sample_processed_data/ has a small sample of the Wizard of Wikipedia dataset files after preprocessing
• tools/ has the evaluation scripts

Requirements

1. The implementation is based on Python 3.x. To install the dependencies used, run:

$ pip install -r requirements.txt

2. Save the pretrained Glove embeddings (glove.840B.300d.txt).

Trained Models

Our trained models on Wizard of Wikipedia and CMU_DoG can be downloaded from here:

https://drive.google.com/drive/folders/1Syn_Q3utg83xgXqCdGGBlKQDChPGDg2n?usp=sharing

Training

For training CNTF, run the following command. All available parameters and flags have been described in main.py. Please refer to paper appendix for further details.

$ CUDA_VISIBLE_DEVICES=x python main.py --data_dir ./sample_processed_data/wiz/ --save_dir ./sample_model/wiz --embed_size 300 --embed_file <path to Glove> --batch_size 8 --vocab_type dlg-seen --bert_config ./source/model/wiz_config.json --num_epochs 30 --log_steps 2

Testing

For testing CNTF, run the following command:

$ CUDA_VISIBLE_DEVICES=x python main.py --data_dir ./sample_processed_data/wiz/ --save_dir ./sample_model/wiz --vocab_type dlg-seen --bert_config ./source/model/wiz_config.json --batch_size 32 --output_dir ./outputs/wiz --test

Evaluation

To evaluate CNTF, run the following command:

$ cd tools/
$ python eval.py --eval_dir ../outputs/wiz --pred_file pred --ref_file tgt

This will generate result.txt which will have all the stated metrics.

To get the predicted responses of the model and gold responses in csv format, run:

$ python read.py --dir ../outputs/wiz

This will generate the required output.csv